Predição dos teores de argila e carbono orgânico do solo integrando dados de sensoriamento remoto

• Ano de defesa: 2026
• Autor(a) principal: Rech, Luís Felipe
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; UFSM; Programa de Pós-Graduação em Ciência do Solo; Centro de Ciências Rurais
• Palavras-chave em Português: Pedometria; Algoritmos de machine learning; Modelagem preditiva; Dados espectrais; Atributos do terreno; Pedometrics; Machine learning algorithms; Predictive modeling; Spectral data; Terrain attributes; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA::CIENCIA DO SOLO
• Link de acesso: http://repositorio.ufsm.br/handle/1/38481

The prediction of soil properties at the spatial scale still faces limitations related to the integration of multiple data sources and the representation of environmental variability. The clay content and soil organic carbon (SOC) are key soil properties that influence biological, physical, and chemical processes. Traditional laboratory methods are time-consuming and require chemical reagents, whereas sensing-based techniques combined with machine learning algorithms enable rapid and non-destructive predictions. The integration of multiple data sources represents a promising approach to improve predictive models, as spectral data respond to soil physicochemical characteristics, magnetic data act as indirect indicators of mineralogy, and topographic covariates represent spatial variability across the landscape. In this context, the study was structured into two experiments: Study 1 evaluated clay content prediction, and Study 2 assessed SOC prediction, both based on the integration of spectral, magnetic, and topographic covariates derived from digital terrain models (DTMs) at multiple spatial resolutions (5, 10, and 30 m) and soil depths (0–5, 5–15, 15–30, 30–60, and 60–100 cm). The machine learning algorithms used to develop the predictive models were Cubist (CUB), Partial Least Squares Regression (PLSR), and Random Forest (RF). In Study 1, the CUB algorithm showed superior predictive performance at all depths, with spectral variables being the most important, followed by magnetic and topographic variables. In Study 2, CUB and PLSR achieved the best performance, with data integration improving SOC predictions up to 30 cm depth. DTM spatial resolution influenced model performance, with 10 m resolution being more relevant in surface layers and 30 m in the intermediate layer. The results highlight the potential of integrating spectral, magnetic, and topographic data for soil property prediction.